Socially Driven Web Sites for the Masses

Transcription

1 Socially Driven Web Sites for the Masses Frank Uyeda Diwaker Gupta, Amin Vahdat, George Varghese University of California, San Diego

2 Grass roots communides wish to have websites that allow them to submit and flexibly search for data. These communides require tools that are simpler than those currently available (e.g. Apache, PHP, MySQL). To build such tools we need models for: data objects, page layout, and, most importantly, search. We have instandated these model in an easy to use language called GrassRoots, and compiler called GR.

3 Web ApplicaDon Pre packaged Custom GrassRoots ConfiguraDon Wikis, Blogs, CMS (WordPress, Joomla) PresentaDon (Dreamweaver) ApplicaDon Logic (PHP, Ruby, Python) Web Frameworks (Zend, Rails, Django) PresentaDon ApplicaDon Model ApplicaDon Logic (Auto Generated) Database ConfiguraDon (Pre specified) Database ConfiguraDon Database (MySQL) Database ConfiguraDon (Auto Generated)

4 MoDvaDon Modeling Results

5 Big social networks internadonal phenomenon. North America: Facebook (250M) South America / India: Orkut (67M users) Not just social networks YouTube (video sharing), Digg (social bookmarking) Growing interest in smaller sites specialized by industry, enterprise and communides. Wellpoint (insurance) Cisco (company specific)

6 Lots of people! Example: Physics Researchers Want to share data sets, tag interesdng features Example: Digital ArDsts Want to share data visualizadon programs & collaborate Example: Local Parents & Baby sigers Want job posdngs, referral network Require database & applicadon logic. These communides lack resources and experdse. Need cheap, easy to use tools!

7 Difficult to prototype new ideas Not clear which web development framework will work best Require knowledge of database schemas, web programming languages, design techniques Significant Dme and experdse needed to develop an operadonal site Large, complex code base IntegraDon with user management, access control and web API s Engineer for security and privacy General techniques to scale are unknown Performance tuning is black magic Hire a consultant Addressed in this talk

8 Site Objects Search Flickr Images Keyword, Tags, Comparison (geo tags) YouTube Video Keyword, Tags Last.fm Audio Tags, Structural Del.icio.us URLs Tags Digg URLs Taxonomy, Keyword Craigslist LisDngs (Image + Text) Taxonomy, Keyword, Comparison Wikipedia ArDcles (Text) Keyword, Structural Facebook User Profile (Image + Text) Structural, Tags

9 Users create and upload content Content organized/ranked based user input Search based on: Associated keywords (e.g., Del.icio.us tags) Structural reladonships (e.g., friends in Facebook) Taxonomy / Hierarchy (e.g., categories in Digg) Comparison / Proximity (e.g., geo tagging in Flickr)

10 Model: Create an abstract model for community driven web sites. Specify: Allow developers to express an abstract site model in the GrassRoots language. Compile: The GrassRoots compiler generates web code and configures storage.

11 MoDvaDon Modeling Results

12 Insight 1 (Layout): Pages are composed of panes that are populated by search results. Insight 2 (NavigaDon): All navigadon is search. Insight 3 (Search): Graph based search with agribute filtering covers exisdng social search mechanisms.

13 GrassRoots objects: High level types (e.g., video, image, text), Composite types like C structs Built in agributes: taggable, commentable RelaDonships as Graphs General Graph (e.g. friends in Facebook) Directed Graph (e.g. YouTube Subscribers) Hierarchy / Tree (e.g. Craigslist categories)

14 What is? Community photo sharing website Users associate with each other Images organized using: Keyword tags User photo sets Group pools

15 StaDc Pane Search Pane Picture Summary Pane Set Summary Pane

16 A page is composed of one or more panes Panes are populated by embedded searches Pane A region within the Page Handles the input and output of a pardcular data collecdon, or displays stadc content Defined once, and reused across many pages Pane aesthedcs customized with CSS

17 User specified Search Pre specified Search

18 NavigaDon associates clicking on a data object to a page and search parameters. Pages embed searches. the linkto keyword provides parameters to searches. Syntax: [object in pane] > linkto [page]( [params] ); Example: Clicking on a user s name displays all pictures owned by that user. user > linkto all_users_pictures( user );

19 Hierarchy Agribute Filter Ordering

20 SELECT <collecdon> [FROM <structural reladon>] [WHERE <filter condidon> ] [ORDER BY <ranking funcdon> ] Structural reladons: Graphs: neighbor, Tree: subtree, parent, children Filter condidons: matches, contains, greater than, between, within distance, tagged by Ranking: CombinaDon of agributes or graph properdes (e.g. node degree)

21 MoDvaDon Modeling Results

22 Claim: Small changes to the specificadon provide important features at low cost. Example: Flickr tags photos with keywords Facebook tags photos with Users How do we change our Flickr specificadon to incorporate User tagging?

23

24 COMPOSITE Picture { IMAGE pic; TEXT pic_dtle; TEXT pic_descripdon; } (taggable, taggable by USER, commentable); PAGE pic_detail( Picture p ) { Detail(Picture) main : LOOKUP Picture p; } Detail Pane : Picture{ _owner > linkto user_profile( _owner ); _tag > linkto tag_result( _tag ); _tag.user > linkto user_profile( _tag.user ); "add tag" > linkto add_tag( this ); "add to set" > linkto add_to_set( this ); "add to group" > linkto add_to_group( this ); }

25 User Tags

26 SpecificaDon Parser Abstract Model Compiler Database Planner Server Scripts Page Generator Database Schema HTML Pages Web Server Implemented in Java (~15K lines of code) Page generadon in various languages. Currently supports PHP Database

27 Implemented Flickr like site using Rails plug ins & GR Code Complexity 50 lines of code across 19 files (vs. 180 lines in 1 file) Picture retrieval throughput: Grassroots gives 2x max throughput. Grassroots only generates necessary code Ruby has large call tree. Tag search throughput 20 most recent 500:1 performance difference, favoring GrassRoots. Suspect poor SQL queries & failure to parallelize.

28 Need beger tools for construcdng social sites. Leverage the commonalides among sites. We provide abstracdons to ease development. Pages are composed of panes All navigadon is search Graph based search with agribute filtering AbstracDons provide flexibility & opportunity for opdmizadon.